ISSN 1360-1725

UMIST
The Matrix Computation Toolbox for MATLAB
(Version 1.0)
N. J. Higham
Numerical Analysis Report No. 410
August 2002
Manchester Centre for Computational Mathematics
Numerical Analysis Reports
DEPARTMENTS OF MATHEMATICS
Reports available from:
Department of Mathematics
University of Manchester
Manchester M13 9PL
England
And over the World-Wide Web from URLs
http://www.ma.man.ac.uk/MCCM
http://www.ma.man.ac.uk/~nareports
The Matrix Computation Toolbox for MATLAB
(Version 1.0)
Nicholas J. Higham
1
August 23, 2002
1
Department of Mathematics, University of Manchester, Manchester, M13 9PL, Eng-
land (higham@ma.man.ac.uk, http://www.ma.man.ac.uk/~higham/).
Contents
1 Introduction 2
1.1 Citing the Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Installation 3
3 Release History 4
4 Quick Reference Tables 5
5 Visualization 7
6 Direct Search Optimization 13
7 Test Matrices 16
Bibliography 19
1
Chapter 1
Introduction
The Matrix Computation Toolbox is a collection of MATLAB M-les containing
functions for constructing test matrices, computing matrix factorizations, visual-
izing matrices, and carrying out direct search optimization. Various other miscel-
laneous functions are also included.
The toolbox was developed in conjunction with the book Accuracy and Stability
of Numerical Algorithms [8, :oo:]. That book is the primary documentation
for the toolbox: it describes much of the underlying mathematics and many of
the algorithms and matrices (it also describes many of the matrices provided by
MATLABs gallery function).
The toolbox is distributed under the terms of the GNU General Public License
(http://www.gnu.org/copyleft/gpl.html, version 2 of the License, or any later
version) as published by the Free Software Foundation.
The toolbox has been tested under MATLAB 6.1 (R12.1) and MATLAB 6.5
(R13).
The M-les in the toolbox are self-documenting and so more detailed documen-
tation than is provided here can be obtained on-line by typing help M-file_name.
For information about MATLAB functions, and in particular the test matrices in
MATLAB, see MATLAB Guide [2, :ooo].
This document describes version 1.0 of the toolbox, dated August 23, 2002.
1.1. Citing the Toolbox
Please cite the toolbox as follows:
N. J. Higham. The Matrix Computation Toolbox. http://www.ma.
man.ac.uk/~higham/mctoolbox.
A BibT
E
X bib entry is available from the Web page shown.
2
Chapter 2
Installation
The Matrix Computation Toolbox is available from
http://www.ma.man.ac.uk/~higham/mctoolbox
It is provided as a zip le that can be uncompressed with any zip le utility. It is
also available from the MATLAB File Exchange at The MathWorks web site
http://www.mathworks.com
It is recommended that the toolbox be installed into a directory matrixcomp.
To try out the toolbox from within MATLAB, change to the matrixcomp directory
and run the demonstration script mctdemo (by typing mctdemo). For serious use it
is best to put the matrixcomp directory on the MATLAB path, so that the M-les
can be called from other directories.
This document is mctoolbox.pdf within the zip le.
3
Chapter 3
Release History
This toolbox supersedes the Test Matrix Toolbox [7, 1]. Most of the test
matrices in that toolbox have been incorporated into MATLAB in the gallery
function. The new toolbox incorporates some of the other routines in the Test
Matrix Toolbox (in some cases renamed) and adds several new ones.
The rst release of the Test Matrix Toolbox (version 1.0) was described in a
technical report [3, 18]. The collection was subsequently published as ACM
Algorithm 694 [4, 11]. Ensuing releases were version 2.0 [6, 1] and version
3.0 [7, 1] (the nal release).
4
Chapter 4
Quick Reference Tables
This section contains quick reference tables for the Matrix Computation Toolbox.
All the M-les in the toolbox are listed by category, with a short description.
Demonstration
mctdemo Demonstration of Matrix Computation Toolbox.
Test Matrices
augment Augmented system matrix.
gfpp Matrix giving maximal growth factor for Gaussian elimination
with partial pivoting.
makejcf A matrix with specied Jordan canonical form.
rschur An upper quasi-triangular matrix.
vand Vandermonde matrix.
vecperm Vec-permutation matrix.
Factorizations and Decompositions
cholp Cholesky factorization with complete pivoting of a positive
semidenite matrix.
cod Complete orthogonal decomposition.
gep Gaussian elimination with pivoting: none, complete, partial,
or rook.
gj GaussJordan elimination with no pivoting or partial pivoting
to solve r = /.
gqr Generalized QR factorization.
gs c Classical GramSchmidt QR factorization.
gs m Modied GramSchmidt QR factorization.
ldlt skew Block LDL
T
factorization for a skew-symmetric matrix.
ldlt symm Block LDL
T
factorization with partial pivoting or rook pivot-
ing for a symmetric indenite matrix.
ldlt sytr Block LDL
T
factorization for a symmetric tridiagonal matrix.
matsignt Matrix sign function of a triangular matrix.
poldec Polar decomposition.
signm Matrix sign decomposition.
trap2tri Unitary reduction of trapezoidal matrix to triangular form.
5
Visualization
fv Field of values (or numerical range).
gersh Gershgorin disks.
ps Dot plot of a pseudospectrum.
pscont Contours and colour pictures of pseudospectra.
see Pictures of a matrix.
Direct Search Optimization
adsmax Alternating directions method.
mdsmax Multidirectional search method.
nmsmax NelderMead simplex method.
Miscellaneous
chop Round matrix elements.
cpltaxes Determine suitable axis for plot of complex vector.
dual Dual vector with respect to Holder j-norm.
lse Solve the equality constrained least squares problem.
matrix Matrix Computation Toolbox information and matrix access
by number.
pnorm Estimate of matrix j-norm (1 j ).
rootm 1th root of a matrix.
seqcheb Sequence of points related to Chebyshev polynomials.
seqm Multiplicative sequence.
show Display signs of matrix elements.
skewpart Skew-symmetric (skew-Hermitian) part.
sparsify Randomly set matrix elements to zero.
strassen Strassens fast matrix multiplication algorithm.
strassenw Strassens fast matrix multiplication algorithm (Winograd
variant).
sub Principal submatrix.
symmpart Symmetric (Hermitian) part.
treshape Reshape vector to or from (unit) triangular matrix.
6
Chapter 5
Visualization
The toolbox contains ve functions for visualizing matrices. The functions can
give insight into the properties of a matrix that is not easy to obtain by looking at
the numerical entries. They also provide an easy way to generate pretty pictures!
The function fv plots the eld of values of a square matrix C
nn
(also
called the numerical range), which is the set of all Rayleigh quotients,

r
r

r
: 0 = r C
n

;
the eigenvalues of are plotted as crosses. The eld of values is a convex set
that contains the eigenvalues. It is the convex hull of the eigenvalues when is
a normal matrix. If is Hermitian, the eld of values is just a segment of the
real line. For non-Hermitian the eld of values is usually two-dimensional and
its shape and size gives some feel for the behaviour of the matrix. Trefethen and
Embree [16] note that the eld of values is the largest reasonable answer to the
question Where in C does a matrix live ? and the spectrum is the smallest
reasonable answer.
Some examples of eld of values plots are given in Figure 5.1. The matrix
gallery(circul,...) is normal, hence its eld of values is the convex hull of
the eigenvalues. For an example of how the eld of values gives insight into the
problem of nding a nearest normal matrix see [12, 18]. An excellent reference
for the theory of the eld of values is [9, 18, Chapter 1].
The function gersh plots the Gershgorin disks of C
nn
, which are the n
disks
1
i
= { . C : |. o
ii
|
n

j=1
j=i
|o
ij
| }
in the complex plane. Gershgorins theorem tells us that the eigenvalues of
lie in the union of the disks, and an extension of the theorem states that if /
disks form a connected region that is isolated from the other disks then there
are precisely / eigenvalues in this region. Thus the size of the disks gives a feel
for how nearly diagonal is, and their locations give information on where the
eigenvalues lie in the complex plane. Four examples of Gershgorin disk plots are
given in Figure 5.2; Gershgorins theorem provides nontrivial information only for
the third matrix, ipjfact(8,1).
The functions ps and pscont plot pseudospectra. The c-pseudospectrum of a
7
2 0 2 4
2
0
2
gallery(grcar,20)
10 5 0 5
5
0
5
rschur(16,2))
0 20 40
20
10
0
10
20
gallery(circul,8)
20 15 10 5
10
5
0
5
10
gallery(lesp,8)
Figure 5.1. Fields of values (fv).
40 20 0
20
10
0
10
20
gallery(lesp,12)
5 0 5
5
0
5
gallery(hanowa,10)
0.2 0 0.2 0.4 0.6 0.8
0.5
0
0.5
gallery(ipjfact,8,1)
2 0 2
2
1
0
1
2
gallery(smoke,16,1
Figure 5.2. Gershgorin disks (gersh).
8
1 0 1
1
0.5
0
0.5
1
(0,1/2,0,0,1)
0 1 2
1
0.5
0
0.5
1
(0,1,1,0,1/4), inverse matrix
0 1 2 3
1
0
1
(0,1/2,1,1,1)
0.5 0 0.5 1 1.5
1
0.5
0
0.5
1
(0,1,0,0,1/4)
Figure 5.3. Pseudospectra of 32 32 pentadiagonal Toeplitz matrices,
gallery(toeppen,32,a,b,c,d,e). Shown are the parameters (a,b,c,d,e).
matrix C
nn
is dened, for a given c 0, to be the set

() = { . : . is an eigenvalue of + 1 for some 1 with 1

2
c }.
In other words, it is the set of all complex numbers that are eigenvalues of + 1
for some perturbation 1 of 2-norm at most c. For a normal matrix the c-
pseudospectrum is the union of the balls of radius c around the eigenvalues of .
For nonnormal matrices the c-pseudospectrum can take a wide variety of shapes
and sizes, depending on the matrix and how nonnormal it is. Pseudospectra play
an important role in many numerical problems. For full details see the work of
Trefethenin particular, [1], [14, 1], [15, 1], [16].
The routine ps plots an approximation to the c-pseudospectrum

(), which
it obtains by computing the eigenvalues of a given number of random perturbations
of . The eigenvalues are plotted as crosses and the pseudo-eigenvalues as dots.
Arguments to ps control the number and type of perturbations. Figure 5.3, taken
from [8, :oo:, Fig. 28.3], gives four examples of 10
3
-pseudospectra, all of which
involve the pentadiagonal Toeplitz matrix gallery(pentoep...).
Another characterization of

(), in terms of the resolvent (.1 )

1
, is

() = { . : (.1 )
1

2
c
1
}.
An alternative way of viewing the pseudospectrum is to plot the function
)(.) = (.1 )
1

1
2
=
min
(.1 )
over the complex plane, where
min
denotes the smallest singular value [14, 1999].
The routine pscont plots log
10
)(.)
1
and oers several ways to view the surface:
9
by its contour lines alone, or as a coloured surface plot in two or three dimensions,
with or without contour lines. (The two-dimensional plot is the view from directly
above the surface.) Two dierent pscont views of the pseudospectra of the trian-
gular matrix gallery(triw,11) are given in Figures 5.4 and 5.5. Since all the
eigenvalues of this matrix are equal to 1, there is a single point where the resolvent
is unbounded in normthis is the bottomless pit in the pictures. The spike in
Figure 5.5 should be innitely deep; since pscont evaluates )(.) on a nite grid,
the spike has a nite depth dependent on the grid spacing. Also because of the
grid spacing chosen, the contours are a little jagged. Various aspects of the plots
can be changed from the MATLAB command line upon return from pscont; for
example, the colour map (colormap), the shading (shading), and the viewing
angle (view). For Figure 5.4 we set shading interp and colormap copper.
Both pseudospectrum functions are computationally intensive, so the defaults
for the arguments are chosen to produce a result in a reasonable time; for plots
that reveal reasonable detail it is usually necessary to override the defaults.
Note that pscont is not written to be ecient. A much more ecient and
versatile tool is the GUI eigtool by Tom Wright [17].
The function see displays a gure with four subplots in the format
mesh(A) semilogy(svd(A))
ps(A) fv(A)
An example for the gallery(chebvand,...) matrix is given in Figure 5.6.
MATLABs mesh command plots a three-dimensional, coloured, wire-frame sur-
face, by regarding the entries of a matrix as specifying heights above a plane.
We use axis(ij), so that the coordinate system for the plot matches the (i, ,)
matrix element numbering. semilogy(svd(A)) plots the singular values of (or-
dered in decreasing size) on a logarithmic scale; the singular values are denoted by
circles, which are joined by a solid line to emphasize the shape of the distribution.
For a sparse MATLAB matrix, see simply displays a spy plot, which shows
the sparsity pattern of the matrix. The user could, of course, try see(full(A))
for a sparse matrix, but for large dimensions the storage and time required would
be prohibitive.
10
0.5 0 0.5 1 1.5
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Figure 5.4. pscont(gallery(triw,11), 0, 30, [-0.5 1.5 -1 1]).
2
0
2
2
0
2
10
8
6
4
2
0
Figure 5.5. pscont(gallery(triw,11), 2, 15, [-2 2 -2 2]).
11
0
5
10
0
5
10
1
0
1
0 5 10
10
5
10
0
10
5
2 0 2
2
1
0
1
2
2 0 2
2
0
2
Figure 5.6. see(gallery(chebvand,8)).
12
Chapter 6
Direct Search Optimization
The toolbox contains three multivariate direct search maximization functions:
mdsmax, adsmax and nmsmax. The functions are competitors to fminsearch, which
is supplied with MATLAB (but fmins minimizes rather than maximizes).
mdsmax, adsmax and nmsmax are direct search methods (as is fminsearch),
that is, they attempt to maximize a real function ) of a vector argument r using
function values only. mdsmax uses the multidirectional search method [10, 1,
8.2], [13, 11], adsmax uses the method of alternating directions, and nmsmax
(like fminsearch) uses the NelderMead simplex method [11, 16], [10, 1,
8.1]. In general, mdsmax and nmsmax can be expected to perform better than
adsmax since they use a more sophisticated method.
These routines were developed during the work described in [5].
It is important to note that these routines, like fminsearch, are not competitive
with more sophisticated methods such as (quasi-)Newton methods when applied
to smooth problems. They are at their best when applied to non-smooth problems
such as the one in the example below.
The routines are fully documented in their leading comment lines, but it is
appropriate to add here a few comments about the format of the output and the
use of the savit argument.
mdsmax produces output to the screen (this can be suppressed by setting the
input argument stopit(5) = 0). The output is illustrated by
Iter. 10, inner = 2, size = -4, nf = 401, f = 4.7183e+001 (51.0%)
The means that on the tenth iteration, two inner iterations were required, and at
the end of the iteration the simplex edges were 2
4
times the length of those of
the initial simplex. Further, nf is the total number of function evaluations so far,
f is the current highest function value, and the percentage increase in function
value over the tenth iteration is 51%.
The output produced by adsmax is similar to that of mdsmax and is illustrated
by the following extract from the start of the second outer iteration:
Iter 2 (nf = 146)
Comp. = 1, steps = 12, f = 1.5607e+000 (0.4%)
Comp denotes the component of r being varied on the current stage and steps is
the number of steps in the crude line search for this stage.
The output from nmsmax is also similar to that from mdsmax, but only iterations
on which an increase in the function value is achieved are reported.
13
In all three routines, if a non-empty fourth input argument string savit is
present then at the end of each iteration the following snapshot is written to
the le specied by savit: the largest function value found so far, fmax, the point
at which it is achieved, x, and the total number of function evaluations, nf. This
option enables the user to abort an optimization, load and examine x, fmax and
nf using MATLABs load command, and then possibly restart the optimization
at x.
One further point worth mentioning is that mdsmax, adsmax and nmsmax always
call the function f to be maximized with an argument of the same dimensions (or
shape) as the starting value x0. Similarly, the output argument x and the variable
x in savit saves have the same shape as x0. This feature is very convenient when
f is a function of a matrix, as in the example below.
To illustrate, here is an example of how to use mdsmax to search for a matrix
for which rcond(A)) overestimates the reciprocal of the exact 1-norm condition
number of the matrix A (rcond is MATLABs built-in condition estimator). The
code
frcond = inline(cond(A,1)*rcond(A));
A = hilb(5); % Starting matrix.
B = mdsmax(@frcond, A)
C = nmsmax(@frcond, B)
uses the output from mdsmax as starting matrix for nmsmax (a useful technique),
and produces the output
f(x0) = 1.0000e+000
Iter. 1, inner = 0, size = 0, nf = 26, f = 1.0000e+000 (0.0%)
Iter. 2, inner = 1, size = -1, nf = 76, f = 1.0000e+000 (0.0%)
Iter. 3, inner = 2, size = -3, nf = 176, f = 1.0066e+000 (0.7%)
Iter. 4, inner = 1, size = -3, nf = 226, f = 1.3521e+000 (34.3%)
Iter. 5, inner = 1, size = -3, nf = 276, f = 1.5476e+000 (14.5%)
Iter. 6, inner = 2, size = -5, nf = 376, f = 1.5697e+000 (1.4%)
Iter. 7, inner = 1, size = -4, nf = 426, f = 1.7432e+000 (11.1%)
Iter. 8, inner = 1, size = -3, nf = 476, f = 1.9645e+000 (12.7%)
Iter. 9, inner = 1, size = -3, nf = 526, f = 3.8308e+000 (95.0%)
Iter. 10, inner = 1, size = -3, nf = 576, f = 4.0586e+000 (5.9%)
Iter. 11, inner = 1, size = -4, nf = 626, f = 4.6397e+000 (14.3%)
Iter. 12, inner = 1, size = -5, nf = 676, f = 4.7817e+000 (3.1%)
Iter. 13, inner = 2, size = -7, nf = 776, f = 4.8090e+000 (0.6%)
Simplex size 8.5236e-004 <= 1.0000e-003...quitting
B =
1.1464 0.5635 0.3913 -0.0014 0.2580
0.5580 0.4576 0.3080 0.2580 0.2246
0.6565 0.3080 0.2580 0.2246 0.2008
0.3245 0.2580 0.2246 0.4218 0.1830
0.2580 0.2246 0.2008 0.1830 0.1691
f(x0) = 4.8090e+000
Iter. 190, how = reflect, nf = 475, f = 4.8357e+000 (0.6%)
Iter. 199, how = reflect, nf = 486, f = 4.8361e+000 (0.0%)
14
Iter. 200, how = expand, nf = 488, f = 4.9610e+000 (2.6%)
Iter. 223, how = reflect, nf = 520, f = 4.9650e+000 (0.1%)
Iter. 224, how = reflect, nf = 522, f = 4.9749e+000 (0.2%)
Iter. 225, how = expand, nf = 524, f = 5.0839e+000 (2.2%)
Iter. 239, how = expand, nf = 542, f = 5.1866e+000 (2.0%)
Iter. 249, how = reflect, nf = 554, f = 5.1913e+000 (0.1%)
Iter. 250, how = reflect, nf = 556, f = 5.2090e+000 (0.3%)
Iter. 251, how = reflect, nf = 558, f = 5.2163e+000 (0.1%)
Iter. 252, how = reflect, nf = 560, f = 5.2330e+000 (0.3%)
Iter. 253, how = reflect, nf = 562, f = 5.2464e+000 (0.3%)
Iter. 254, how = reflect, nf = 564, f = 5.2606e+000 (0.3%)
Iter. 256, how = expand, nf = 567, f = 5.3933e+000 (2.5%)
Iter. 268, how = expand, nf = 581, f = 5.5486e+000 (2.9%)
Iter. 286, how = expand, nf = 604, f = 5.7267e+000 (3.2%)
Iter. 302, how = expand, nf = 624, f = 5.8959e+000 (3.0%)
Iter. 310, how = expand, nf = 633, f = 6.1229e+000 (3.9%)
... (output omitted)
Iter. 927, how = reflect, nf = 1430, f = 2.6536e+002 (4.5%)
Iter. 938, how = reflect, nf = 1447, f = 2.7461e+002 (3.5%)
Iter. 944, how = reflect, nf = 1457, f = 2.7752e+002 (1.1%)
Iter. 962, how = reflect, nf = 1487, f = 2.8312e+002 (2.0%)
Simplex size 8.5771e-004 <= 1.0000e-003...quitting
C =
1.1571 0.5621 0.3576 -0.0102 0.2625
0.5275 0.5064 0.3115 0.2837 0.2286
0.6478 0.3180 0.2752 0.2533 0.2099
0.3254 0.2455 0.2418 0.4101 0.1869
0.2490 0.2156 0.2152 0.1755 0.1698
For more on the use of direct search in automatic error analysis see [8, :oo:,
Ch. 26].
15
Chapter 7
Test Matrices
The function matrix provides easy access to most of the test matrices in MATLAB
and those in the Matrix Computation Toolbox. With no arguments the matrix
function lists the available matrices by number:
>> matrix
1: cauchy 12: frank 23: lesp 34: randsvd 45: rand
2: chebspec 13: gearmat 24: lotkin 35: redheff 46: randn
3: chebvand 14: grcar 25: minij 36: riemann 47: augment
4: chow 15: invhess 26: moler 37: ris 48: gfpp
5: circul 16: invol 27: orthog 38: smoke 49: magic
6: clement 17: ipjfact 28: parter 39: toeppd 50: makejcf
7: condex 18: jordbloc 29: pei 40: triw 51: rschur
8: cycol 19: kahan 30: prolate 41: hilb 52: vand
9: dramadah 20: kms 31: randcolu 42: invhilb
10: fiedler 21: krylov 32: randcorr 43: magic
11: forsythe 22: lehmer 33: rando 44: pascal
Matrices 1 to 46 are from MATLAB
Invoking matrix(k,n,...) produces the n-by-n instance of the kth of these ma-
trices, all of which are full (as opposed to sparse). This provides a convenient way
to run through a batch of test matrices. For example, the code
c = []; j = 1;
% Make experiment repeatable for the random matrices.
randn(seed,1), rand(seed,1)
fprintf(Matrix Ratio\n-------------------\n)
for k=1:matrix(0)
% Double on next line avoids bug in MATLAB 6.5 re. matrix(35).
A = double(matrix(k, 16));
c = [c max(abs(eig(A)))/norm(A)];
name = [matrix(k) ];
fprintf([name(1:8) : %9.1e\n], c(end))
end
16
runs through the set of 52 matrices evaluating the ratio | max
i

i
()|,
2
for
dimension 16, where {
i
} is the set of eigenvalues of . This ratio is known
theoretically to lie between 0 and 1. For only one of the matrices (invol) is the
ratio signicantly less than 1:
Matrix Ratio
-------------------
cauchy : 1.0e+000
chebspec: 1.2e-002
chebvand: 7.4e-001
chow : 3.5e-001
circul : 1.0e+000
clement : 9.4e-001
condex : 1.0e+000
cycol : 4.2e-001
dramadah: 7.5e-001
fiedler : 1.0e+000
forsythe: 3.2e-001
frank : 5.8e-001
gearmat : 9.8e-001
grcar : 7.0e-001
invhess : 9.7e-001
invol : 4.9e-011
ipjfact : 1.0e+000
jordbloc: 5.0e-001
kahan : 3.8e-001
kms : 1.0e+000
krylov : 3.5e-001
lehmer : 1.0e+000
lesp : 7.8e-001
lotkin : 6.4e-001
minij : 1.0e+000
moler : 1.0e+000
orthog : 1.0e+000
parter : 9.5e-001
pei : 1.0e+000
prolate : 1.0e+000
randcolu: 6.8e-001
randcorr: 1.0e+000
rando : 9.7e-001
randsvd : 1.9e-001
redheff : 9.2e-001
riemann : 9.1e-001
ris : 1.0e+000
smoke : 5.2e-001
toeppd : 1.0e+000
triw : 1.1e-001
hilb : 1.0e+000
17
invhilb : 1.0e+000
magic : 1.0e+000
pascal : 1.0e+000
rand : 9.7e-001
randn : 5.5e-001
augment : 1.0e+000
gfpp : 5.0e-001
magic : 1.0e+000
makejcf : 8.6e-002
rschur : 9.2e-001
vand : 2.6e-001
18
Bibliography
[1] Mark Embree and Lloyd N. Trefethen. Pseudospectra gateway. http://www.comlab.
ox.ac.uk/pseudospectra/.
[2] Desmond J. Higham and Nicholas J. Higham. MATLAB Guide. Society for Indus-
trial and Applied Mathematics, Philadelphia, PA, USA, 2000. xxii+283 pp. ISBN
0-89871-516-4.
[3] Nicholas J. Higham. A collection of test matrices in MATLAB. Numerical Analysis
Report No. 172, University of Manchester, Manchester, England, July 1989.
[4] Nicholas J. Higham. Algorithm 694: A collection of test matrices in MATLAB.
ACM Trans. Math. Software, 17(3):289305, September 1991.
[5] Nicholas J. Higham. Optimization by direct search in matrix computations. SIAM
J. Matrix Anal. Appl., 14(2):317333, April 1993.
[6] Nicholas J. Higham. The Test Matrix Toolbox for MATLAB. Numerical Analysis
Report No. 237, Manchester Centre for Computational Mathematics, Manchester,
England, December 1993. 76 pp.
[7] Nicholas J. Higham. The Test Matrix Toolbox for MATLAB (version 3.0). Numer-
ical Analysis Report No. 276, Manchester Centre for Computational Mathematics,
Manchester, England, September 1995. 70 pp.
[8] Nicholas J. Higham. Accuracy and Stability of Numerical Algorithms. Second edi-
tion, Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 2002.
xxx+680 pp. ISBN 0-89871-521-0.
[9] Roger A. Horn and Charles R. Johnson. Topics in Matrix Analysis. Cambridge
University Press, 1991. viii+607 pp. ISBN 0-521-30587-X.
[10] C. T. Kelley. Iterative Methods for Optimization. Society for Industrial and Applied
Mathematics, Philadelphia, PA, USA, 1999. xv+180 pp. ISBN 0-89871-433-8.
[11] J. A. Nelder and R. Mead. A simplex method for function minimization. Comput.
J., 7:308313, 1965.
[12] Axel Ruhe. Closest normal matrix nally found! BIT, 27:585598, 1987.
[13] Virginia J. Torczon. On the convergence of the multidirectional search algorithm.
SIAM J. Optim., 1(1):123145, 1991.
[14] Lloyd N. Trefethen. Computation of pseudospectra. Acta Numerica, 8:247295,
1999.
[15] Lloyd N. Trefethen. Spectra and pseudospectra. In The Graduate Students Guide
to Numerical Analysis 98, Mark Ainsworth, Jeremy Levesley, and Marco Marletta,
editors, Springer-Verlag, Berlin, 1999, pages 217250.
[16] Lloyd N. Trefethen and Mark Embree. Spectra and Pseudospectra: The Behavior of
Non-Normal Matrices and Operators. Book in preparation.
[17] Thomas G. Wright. Eigtool. http://www.comlab.ox.ac.uk/pseudospectra/
eigtool/.
19